Google DeepMind released EmbeddingGemma 2 on October 6, 2026 as an open-weight, multimodal successor to its text-only EmbeddingGemma. Built on Gemma 4, it is designed to make different kinds of content searchable in one shared embedding space, including on a device rather than through a cloud service. 12

• License and size: The new model has 740 million parameters and uses the Apache 2.0 license. The original, released in September 2025, had 308 million parameters and was based on Gemma 3. 134
• Media and architecture: Instead of embedding only text, EmbeddingGemma 2 handles text and code, images, video, audio, and combinations of them. Its modular encoders map those inputs into a shared 768-dimensional vector space, so an application can compare, for example, a text query with visual or audio content without chaining separate embedding models. Developers can load the encoders they need, with configurations ranging from 270 million to 740 million parameters. 15
• Context and vector size: Its reported 8,000-token context is roughly four times the original’s 2,000-token context. Both support smaller output vectors rather than requiring all 768 dimensions; the original explicitly offered 128–768 dimensions. 46
• Benchmarks and memory: Google describes the new model as best-in-class for its size, but the evidence available here is insufficient to give a defensible numerical benchmark comparison with the original. Reported quantized on-device memory use is about 191 MB for the text-only configuration and 567 MB for the full multimodal configuration; these are configuration-specific figures, not a guarantee for every device. 267
• Uses and limitations: The expansion enables local semantic search across notes, photos, recordings and video, as well as multimodal retrieval for applications such as RAG. A shared embedding does not itself establish that retrieved material is safe, accurate or appropriate: applications still need their own handling for sensitive content and mistaken matches. 28
• Availability: The weights are currently listed through Google’s developer materials and its Hugging Face model page. Insufficient evidence from the sources reviewed to state a firm date for any additional planned distribution or integrations. 158

Google DeepMind released EmbeddingGemma 2 on October 6, 2026 as an open-weight, multimodal successor to its text-only EmbeddingGemma. Built on Gemma 4, it is designed to make different kinds of content searchable in one shared embedding space, including on a device rather than through a cloud service. 12

• License and size: The new model has 740 million parameters and uses the Apache 2.0 license. The original, released in September 2025, had 308 million parameters and was based on Gemma 3. 134
• Media and architecture: Instead of embedding only text, EmbeddingGemma 2 handles text and code, images, video, audio, and combinations of them. Its modular encoders map those inputs into a shared 768-dimensional vector space, so an application can compare, for example, a text query with visual or audio content without chaining separate embedding models. Developers can load the encoders they need, with configurations ranging from 270 million to 740 million parameters. 15
• Context and vector size: Its reported 8,000-token context is roughly four times the original’s 2,000-token context. Both support smaller output vectors rather than requiring all 768 dimensions; the original explicitly offered 128–768 dimensions. 46
• Benchmarks and memory: Google describes the new model as best-in-class for its size, but the evidence available here is insufficient to give a defensible numerical benchmark comparison with the original. Reported quantized on-device memory use is about 191 MB for the text-only configuration and 567 MB for the full multimodal configuration; these are configuration-specific figures, not a guarantee for every device. 267
• Uses and limitations: The expansion enables local semantic search across notes, photos, recordings and video, as well as multimodal retrieval for applications such as RAG. A shared embedding does not itself establish that retrieved material is safe, accurate or appropriate: applications still need their own handling for sensitive content and mistaken matches. 28
• Availability: The weights are currently listed through Google’s developer materials and its Hugging Face model page. Insufficient evidence from the sources reviewed to state a firm date for any additional planned distribution or integrations. 158